sparkyniner/netryx-astra-v2-geolocation-toolThe world's most sophisticated street level image geolocation software
State-of-the-art AI geolocation from a single image.
The Idea • How It Works • Getting Started • Community Hub • Installation
You have a photograph. Maybe it's a screenshot from a video. Maybe it's a cropped, blurry phone photo someone posted online. Maybe it shows just a storefront, a stretch of road, or the corner of a building. You want to know exactly where it was taken.
Netryx Astra V2 answers that question.
It's an open-source geolocation system that takes a single image and finds the precise GPS coordinates by matching it against a database of street-view panoramas. Upload your photo, and within minutes it tells you the street, the city, the coordinates — down to a few meters.
What makes V2 different from the original Netryx (and from other tools out there) is the matching pipeline. We rebuilt everything from the ground up using two models that didn't exist when we started this project:
MegaLoc (CVPR 2025) — the most accurate image retrieval model for place recognition, trained across six datasets covering indoor, outdoor, day, night, and seasonal variations. It finds the right neighborhood.
MASt3R (ECCV 2024) — a 3D-aware dense matcher that understands the geometry of scenes, not just pixel patterns. It confirms the exact location, even from partial or heavily cropped photos that would break traditional matchers.
The result is a three-step pipeline that's both simpler and more accurate than the nine-stage system it replaced.
The original Netryx used CosPlace for retrieval and a stack of DISK + LightGlue + LoFTR + RANSAC + descriptor hopping + neighborhood expansion for verification. It worked, but it was fragile — lots of heuristics layered on top of each other, each one a workaround for a limitation in the previous stage.
V2 threw all of that away. Here's what replaced what:
| V1 (Original) | V2 (Astra) | |
|---|---|---|
| Finding candidates | CosPlace (ResNet-50, 512-dim) | MegaLoc (DINOv2 ViT-B/14, 8448-dim → PCA 1024) |
| Confirming matches | DISK + LightGlue + RANSAC | MASt3R dense 3D matching |
| Handling edge cases | LoFTR fallback, descriptor hopping, neighborhood expansion, Ultra Mode | Spatial consensus — that's it |
| Total pipeline stages | 9+ | 3 |
| Partial image matching | Weak — sparse keypoints fail on small overlaps | Strong — MASt3R finds dense correspondences in tiny regions |
| Sharing indexes | Not possible | Community Hub via Hugging Face + offline .netryx bundles |
The simplification isn't just aesthetic. Fewer stages means fewer places for things to go wrong, faster searches, and code that's actually maintainable.
The pipeline has three stages. That's not an oversimplification — it's genuinely just three stages.
Query Image
│
▼
┌─────────────┐
│ MegaLoc │ "Where in the city could this be?"
│ Retrieval │
└─────┬───────┘
│ Top 500 candidates
▼
┌─────────────┐
│ MASt3R │ "Is this actually the same place?"
│ Matching │
└─────┬───────┘
│ Scored candidates
▼
┌─────────────┐
│ Spatial │ "Which cluster of matches is most trustworthy?"
│ Consensus │
└─────┬───────┘
│
▼
📍 GPS Coordinates
Your query image gets converted into a compact descriptor — a 8448-dimensional vector that captures the visual essence of the scene. This gets PCA-reduced to 1024 dimensions, then compared against every indexed location via dot-product similarity.
We also extract a descriptor for a slightly zoomed-in center crop and for a horizontally flipped version of the query, then merge the results. This handles cases where the query is at a different zoom level or facing the opposite direction from the indexed view.
The output is the top 500 candidate locations from the index, ranked by visual similarity.
MegaLoc is from Gabriele Berton's lab (the same group that made CosPlace and EigenPlaces). It's the latest in their line of work, trained on SF-XL, GSV-Cities, MSLS, and landmark retrieval data simultaneously. No other retrieval model consistently beats it across every benchmark — indoor, outdoor, urban, rural, day, night.
For each of those 500 candidates, we download the corresponding street-view panorama, crop it at the indexed heading angle, and run MASt3R to find dense pixel correspondences between the query and the crop.
This is where the magic happens for difficult queries. Traditional matchers like SuperPoint + LightGlue extract maybe 500-2000 sparse keypoints and try to match them. If your query image only overlaps 20% with the database image, there might only be 50 co-visible keypoints — not enough for a reliable match.
MASt3R works completely differently. It treats matching as a 3D reconstruction problem, predicting dense point maps and local feature descriptors for every pixel. Even a small overlapping region produces hundreds of reliable correspondences, because it understands the 3D structure of the scene, not just 2D pixel patterns.
On the Map-free localization benchmark (single reference image, viewpoint changes up to 180°), MASt3R beats previous methods by 30%. That's not an incremental improvement — it's a generational leap.
Here's the problem with just picking the candidate with the highest match score: false positives exist. Two identical chain restaurants 5km apart will both produce high MASt3R scores. A row of Soviet-era apartment blocks all look the same.
Spatial consensus solves this. We divide the search area into ~50-meter grid cells and cluster all the good matches geographically. Each cell gets a score based on the combined evidence from all matches in that cell and its neighbors.
A single outlier with 200 inliers at the wrong location gets outscored by a cluster of 5 matches with 80-150 inliers each at the right location. The winning cluster's best match becomes the final answer.
This is why accuracy holds up even at larger search radii where there are more look-alike locations.
We can observe in this picture, there is absolutely nothing to go on, it is just a small cropped part. Conventional OSINT would absolutely fail here. Yet Netryx Astra geolocated it down to its exact coordinates with no metadta whatsoever or clues beforehand running compltely locally.